1 /*- 2 * Copyright (c) 1991 Regents of the University of California. 3 * All rights reserved. 4 * Copyright (c) 1994 John S. Dyson 5 * All rights reserved. 6 * Copyright (c) 1994 David Greenman 7 * All rights reserved. 8 * Copyright (c) 2005 Yahoo! Technologies Norway AS 9 * All rights reserved. 10 * 11 * This code is derived from software contributed to Berkeley by 12 * The Mach Operating System project at Carnegie-Mellon University. 13 * 14 * Redistribution and use in source and binary forms, with or without 15 * modification, are permitted provided that the following conditions 16 * are met: 17 * 1. Redistributions of source code must retain the above copyright 18 * notice, this list of conditions and the following disclaimer. 19 * 2. Redistributions in binary form must reproduce the above copyright 20 * notice, this list of conditions and the following disclaimer in the 21 * documentation and/or other materials provided with the distribution. 22 * 3. All advertising materials mentioning features or use of this software 23 * must display the following acknowledgement: 24 * This product includes software developed by the University of 25 * California, Berkeley and its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * from: @(#)vm_pageout.c 7.4 (Berkeley) 5/7/91 43 * 44 * 45 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 46 * All rights reserved. 47 * 48 * Authors: Avadis Tevanian, Jr., Michael Wayne Young 49 * 50 * Permission to use, copy, modify and distribute this software and 51 * its documentation is hereby granted, provided that both the copyright 52 * notice and this permission notice appear in all copies of the 53 * software, derivative works or modified versions, and any portions 54 * thereof, and that both notices appear in supporting documentation. 55 * 56 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 57 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 58 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 59 * 60 * Carnegie Mellon requests users of this software to return to 61 * 62 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 63 * School of Computer Science 64 * Carnegie Mellon University 65 * Pittsburgh PA 15213-3890 66 * 67 * any improvements or extensions that they make and grant Carnegie the 68 * rights to redistribute these changes. 69 */ 70 71 /* 72 * The proverbial page-out daemon. 73 */ 74 75 #include <sys/cdefs.h> 76 __FBSDID("$FreeBSD$"); 77 78 #include "opt_vm.h" 79 #include <sys/param.h> 80 #include <sys/systm.h> 81 #include <sys/kernel.h> 82 #include <sys/eventhandler.h> 83 #include <sys/lock.h> 84 #include <sys/mutex.h> 85 #include <sys/proc.h> 86 #include <sys/kthread.h> 87 #include <sys/ktr.h> 88 #include <sys/resourcevar.h> 89 #include <sys/sched.h> 90 #include <sys/signalvar.h> 91 #include <sys/vnode.h> 92 #include <sys/vmmeter.h> 93 #include <sys/sx.h> 94 #include <sys/sysctl.h> 95 96 #include <vm/vm.h> 97 #include <vm/vm_param.h> 98 #include <vm/vm_object.h> 99 #include <vm/vm_page.h> 100 #include <vm/vm_map.h> 101 #include <vm/vm_pageout.h> 102 #include <vm/vm_pager.h> 103 #include <vm/swap_pager.h> 104 #include <vm/vm_extern.h> 105 #include <vm/uma.h> 106 107 #include <machine/mutex.h> 108 109 /* 110 * System initialization 111 */ 112 113 /* the kernel process "vm_pageout"*/ 114 static void vm_pageout(void); 115 static int vm_pageout_clean(vm_page_t); 116 static void vm_pageout_scan(int pass); 117 118 struct proc *pageproc; 119 120 static struct kproc_desc page_kp = { 121 "pagedaemon", 122 vm_pageout, 123 &pageproc 124 }; 125 SYSINIT(pagedaemon, SI_SUB_KTHREAD_PAGE, SI_ORDER_FIRST, kproc_start, &page_kp) 126 127 #if !defined(NO_SWAPPING) 128 /* the kernel process "vm_daemon"*/ 129 static void vm_daemon(void); 130 static struct proc *vmproc; 131 132 static struct kproc_desc vm_kp = { 133 "vmdaemon", 134 vm_daemon, 135 &vmproc 136 }; 137 SYSINIT(vmdaemon, SI_SUB_KTHREAD_VM, SI_ORDER_FIRST, kproc_start, &vm_kp) 138 #endif 139 140 141 int vm_pages_needed; /* Event on which pageout daemon sleeps */ 142 int vm_pageout_deficit; /* Estimated number of pages deficit */ 143 int vm_pageout_pages_needed; /* flag saying that the pageout daemon needs pages */ 144 145 #if !defined(NO_SWAPPING) 146 static int vm_pageout_req_swapout; /* XXX */ 147 static int vm_daemon_needed; 148 #endif 149 static int vm_max_launder = 32; 150 static int vm_pageout_stats_max=0, vm_pageout_stats_interval = 0; 151 static int vm_pageout_full_stats_interval = 0; 152 static int vm_pageout_algorithm=0; 153 static int defer_swap_pageouts=0; 154 static int disable_swap_pageouts=0; 155 156 #if defined(NO_SWAPPING) 157 static int vm_swap_enabled=0; 158 static int vm_swap_idle_enabled=0; 159 #else 160 static int vm_swap_enabled=1; 161 static int vm_swap_idle_enabled=0; 162 #endif 163 164 SYSCTL_INT(_vm, VM_PAGEOUT_ALGORITHM, pageout_algorithm, 165 CTLFLAG_RW, &vm_pageout_algorithm, 0, "LRU page mgmt"); 166 167 SYSCTL_INT(_vm, OID_AUTO, max_launder, 168 CTLFLAG_RW, &vm_max_launder, 0, "Limit dirty flushes in pageout"); 169 170 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_max, 171 CTLFLAG_RW, &vm_pageout_stats_max, 0, "Max pageout stats scan length"); 172 173 SYSCTL_INT(_vm, OID_AUTO, pageout_full_stats_interval, 174 CTLFLAG_RW, &vm_pageout_full_stats_interval, 0, "Interval for full stats scan"); 175 176 SYSCTL_INT(_vm, OID_AUTO, pageout_stats_interval, 177 CTLFLAG_RW, &vm_pageout_stats_interval, 0, "Interval for partial stats scan"); 178 179 #if defined(NO_SWAPPING) 180 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 181 CTLFLAG_RD, &vm_swap_enabled, 0, ""); 182 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 183 CTLFLAG_RD, &vm_swap_idle_enabled, 0, ""); 184 #else 185 SYSCTL_INT(_vm, VM_SWAPPING_ENABLED, swap_enabled, 186 CTLFLAG_RW, &vm_swap_enabled, 0, "Enable entire process swapout"); 187 SYSCTL_INT(_vm, OID_AUTO, swap_idle_enabled, 188 CTLFLAG_RW, &vm_swap_idle_enabled, 0, "Allow swapout on idle criteria"); 189 #endif 190 191 SYSCTL_INT(_vm, OID_AUTO, defer_swapspace_pageouts, 192 CTLFLAG_RW, &defer_swap_pageouts, 0, "Give preference to dirty pages in mem"); 193 194 SYSCTL_INT(_vm, OID_AUTO, disable_swapspace_pageouts, 195 CTLFLAG_RW, &disable_swap_pageouts, 0, "Disallow swapout of dirty pages"); 196 197 static int pageout_lock_miss; 198 SYSCTL_INT(_vm, OID_AUTO, pageout_lock_miss, 199 CTLFLAG_RD, &pageout_lock_miss, 0, "vget() lock misses during pageout"); 200 201 #define VM_PAGEOUT_PAGE_COUNT 16 202 int vm_pageout_page_count = VM_PAGEOUT_PAGE_COUNT; 203 204 int vm_page_max_wired; /* XXX max # of wired pages system-wide */ 205 206 #if !defined(NO_SWAPPING) 207 static void vm_pageout_map_deactivate_pages(vm_map_t, long); 208 static void vm_pageout_object_deactivate_pages(pmap_t, vm_object_t, long); 209 static void vm_req_vmdaemon(void); 210 #endif 211 static void vm_pageout_page_stats(void); 212 213 /* 214 * vm_pageout_fallback_object_lock: 215 * 216 * Lock vm object currently associated with `m'. VM_OBJECT_TRYLOCK is 217 * known to have failed and page queue must be either PQ_ACTIVE or 218 * PQ_INACTIVE. To avoid lock order violation, unlock the page queues 219 * while locking the vm object. Use marker page to detect page queue 220 * changes and maintain notion of next page on page queue. Return 221 * TRUE if no changes were detected, FALSE otherwise. vm object is 222 * locked on return. 223 * 224 * This function depends on both the lock portion of struct vm_object 225 * and normal struct vm_page being type stable. 226 */ 227 static boolean_t 228 vm_pageout_fallback_object_lock(vm_page_t m, vm_page_t *next) 229 { 230 struct vm_page marker; 231 boolean_t unchanged; 232 u_short queue; 233 vm_object_t object; 234 235 /* 236 * Initialize our marker 237 */ 238 bzero(&marker, sizeof(marker)); 239 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 240 marker.queue = m->queue; 241 marker.wire_count = 1; 242 243 queue = m->queue; 244 object = m->object; 245 246 TAILQ_INSERT_AFTER(&vm_page_queues[queue].pl, 247 m, &marker, pageq); 248 vm_page_unlock_queues(); 249 VM_OBJECT_LOCK(object); 250 vm_page_lock_queues(); 251 252 /* Page queue might have changed. */ 253 *next = TAILQ_NEXT(&marker, pageq); 254 unchanged = (m->queue == queue && 255 m->object == object && 256 &marker == TAILQ_NEXT(m, pageq)); 257 TAILQ_REMOVE(&vm_page_queues[queue].pl, 258 &marker, pageq); 259 return (unchanged); 260 } 261 262 /* 263 * vm_pageout_clean: 264 * 265 * Clean the page and remove it from the laundry. 266 * 267 * We set the busy bit to cause potential page faults on this page to 268 * block. Note the careful timing, however, the busy bit isn't set till 269 * late and we cannot do anything that will mess with the page. 270 */ 271 static int 272 vm_pageout_clean(m) 273 vm_page_t m; 274 { 275 vm_object_t object; 276 vm_page_t mc[2*vm_pageout_page_count]; 277 int pageout_count; 278 int ib, is, page_base; 279 vm_pindex_t pindex = m->pindex; 280 281 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 282 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 283 284 /* 285 * It doesn't cost us anything to pageout OBJT_DEFAULT or OBJT_SWAP 286 * with the new swapper, but we could have serious problems paging 287 * out other object types if there is insufficient memory. 288 * 289 * Unfortunately, checking free memory here is far too late, so the 290 * check has been moved up a procedural level. 291 */ 292 293 /* 294 * Don't mess with the page if it's busy, held, or special 295 */ 296 if ((m->hold_count != 0) || 297 ((m->busy != 0) || (m->flags & (PG_BUSY|PG_UNMANAGED)))) { 298 return 0; 299 } 300 301 mc[vm_pageout_page_count] = m; 302 pageout_count = 1; 303 page_base = vm_pageout_page_count; 304 ib = 1; 305 is = 1; 306 307 /* 308 * Scan object for clusterable pages. 309 * 310 * We can cluster ONLY if: ->> the page is NOT 311 * clean, wired, busy, held, or mapped into a 312 * buffer, and one of the following: 313 * 1) The page is inactive, or a seldom used 314 * active page. 315 * -or- 316 * 2) we force the issue. 317 * 318 * During heavy mmap/modification loads the pageout 319 * daemon can really fragment the underlying file 320 * due to flushing pages out of order and not trying 321 * align the clusters (which leave sporatic out-of-order 322 * holes). To solve this problem we do the reverse scan 323 * first and attempt to align our cluster, then do a 324 * forward scan if room remains. 325 */ 326 object = m->object; 327 more: 328 while (ib && pageout_count < vm_pageout_page_count) { 329 vm_page_t p; 330 331 if (ib > pindex) { 332 ib = 0; 333 break; 334 } 335 336 if ((p = vm_page_lookup(object, pindex - ib)) == NULL) { 337 ib = 0; 338 break; 339 } 340 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) || 341 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 342 ib = 0; 343 break; 344 } 345 vm_page_test_dirty(p); 346 if ((p->dirty & p->valid) == 0 || 347 p->queue != PQ_INACTIVE || 348 p->wire_count != 0 || /* may be held by buf cache */ 349 p->hold_count != 0) { /* may be undergoing I/O */ 350 ib = 0; 351 break; 352 } 353 mc[--page_base] = p; 354 ++pageout_count; 355 ++ib; 356 /* 357 * alignment boundry, stop here and switch directions. Do 358 * not clear ib. 359 */ 360 if ((pindex - (ib - 1)) % vm_pageout_page_count == 0) 361 break; 362 } 363 364 while (pageout_count < vm_pageout_page_count && 365 pindex + is < object->size) { 366 vm_page_t p; 367 368 if ((p = vm_page_lookup(object, pindex + is)) == NULL) 369 break; 370 if (VM_PAGE_INQUEUE1(p, PQ_CACHE) || 371 (p->flags & (PG_BUSY|PG_UNMANAGED)) || p->busy) { 372 break; 373 } 374 vm_page_test_dirty(p); 375 if ((p->dirty & p->valid) == 0 || 376 p->queue != PQ_INACTIVE || 377 p->wire_count != 0 || /* may be held by buf cache */ 378 p->hold_count != 0) { /* may be undergoing I/O */ 379 break; 380 } 381 mc[page_base + pageout_count] = p; 382 ++pageout_count; 383 ++is; 384 } 385 386 /* 387 * If we exhausted our forward scan, continue with the reverse scan 388 * when possible, even past a page boundry. This catches boundry 389 * conditions. 390 */ 391 if (ib && pageout_count < vm_pageout_page_count) 392 goto more; 393 394 /* 395 * we allow reads during pageouts... 396 */ 397 return (vm_pageout_flush(&mc[page_base], pageout_count, 0)); 398 } 399 400 /* 401 * vm_pageout_flush() - launder the given pages 402 * 403 * The given pages are laundered. Note that we setup for the start of 404 * I/O ( i.e. busy the page ), mark it read-only, and bump the object 405 * reference count all in here rather then in the parent. If we want 406 * the parent to do more sophisticated things we may have to change 407 * the ordering. 408 */ 409 int 410 vm_pageout_flush(vm_page_t *mc, int count, int flags) 411 { 412 vm_object_t object = mc[0]->object; 413 int pageout_status[count]; 414 int numpagedout = 0; 415 int i; 416 417 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 418 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED); 419 /* 420 * Initiate I/O. Bump the vm_page_t->busy counter and 421 * mark the pages read-only. 422 * 423 * We do not have to fixup the clean/dirty bits here... we can 424 * allow the pager to do it after the I/O completes. 425 * 426 * NOTE! mc[i]->dirty may be partial or fragmented due to an 427 * edge case with file fragments. 428 */ 429 for (i = 0; i < count; i++) { 430 KASSERT(mc[i]->valid == VM_PAGE_BITS_ALL, 431 ("vm_pageout_flush: partially invalid page %p index %d/%d", 432 mc[i], i, count)); 433 vm_page_io_start(mc[i]); 434 pmap_page_protect(mc[i], VM_PROT_READ); 435 } 436 vm_page_unlock_queues(); 437 vm_object_pip_add(object, count); 438 439 vm_pager_put_pages(object, mc, count, 440 (flags | ((object == kernel_object) ? VM_PAGER_PUT_SYNC : 0)), 441 pageout_status); 442 443 vm_page_lock_queues(); 444 for (i = 0; i < count; i++) { 445 vm_page_t mt = mc[i]; 446 447 KASSERT((mt->flags & PG_WRITEABLE) == 0, 448 ("vm_pageout_flush: page %p is not write protected", mt)); 449 switch (pageout_status[i]) { 450 case VM_PAGER_OK: 451 case VM_PAGER_PEND: 452 numpagedout++; 453 break; 454 case VM_PAGER_BAD: 455 /* 456 * Page outside of range of object. Right now we 457 * essentially lose the changes by pretending it 458 * worked. 459 */ 460 pmap_clear_modify(mt); 461 vm_page_undirty(mt); 462 break; 463 case VM_PAGER_ERROR: 464 case VM_PAGER_FAIL: 465 /* 466 * If page couldn't be paged out, then reactivate the 467 * page so it doesn't clog the inactive list. (We 468 * will try paging out it again later). 469 */ 470 vm_page_activate(mt); 471 break; 472 case VM_PAGER_AGAIN: 473 break; 474 } 475 476 /* 477 * If the operation is still going, leave the page busy to 478 * block all other accesses. Also, leave the paging in 479 * progress indicator set so that we don't attempt an object 480 * collapse. 481 */ 482 if (pageout_status[i] != VM_PAGER_PEND) { 483 vm_object_pip_wakeup(object); 484 vm_page_io_finish(mt); 485 if (vm_page_count_severe()) 486 vm_page_try_to_cache(mt); 487 } 488 } 489 return numpagedout; 490 } 491 492 #if !defined(NO_SWAPPING) 493 /* 494 * vm_pageout_object_deactivate_pages 495 * 496 * deactivate enough pages to satisfy the inactive target 497 * requirements or if vm_page_proc_limit is set, then 498 * deactivate all of the pages in the object and its 499 * backing_objects. 500 * 501 * The object and map must be locked. 502 */ 503 static void 504 vm_pageout_object_deactivate_pages(pmap, first_object, desired) 505 pmap_t pmap; 506 vm_object_t first_object; 507 long desired; 508 { 509 vm_object_t backing_object, object; 510 vm_page_t p, next; 511 int actcount, rcount, remove_mode; 512 513 VM_OBJECT_LOCK_ASSERT(first_object, MA_OWNED); 514 if (first_object->type == OBJT_DEVICE || first_object->type == OBJT_PHYS) 515 return; 516 for (object = first_object;; object = backing_object) { 517 if (pmap_resident_count(pmap) <= desired) 518 goto unlock_return; 519 if (object->paging_in_progress) 520 goto unlock_return; 521 522 remove_mode = 0; 523 if (object->shadow_count > 1) 524 remove_mode = 1; 525 /* 526 * scan the objects entire memory queue 527 */ 528 rcount = object->resident_page_count; 529 p = TAILQ_FIRST(&object->memq); 530 vm_page_lock_queues(); 531 while (p && (rcount-- > 0)) { 532 if (pmap_resident_count(pmap) <= desired) { 533 vm_page_unlock_queues(); 534 goto unlock_return; 535 } 536 next = TAILQ_NEXT(p, listq); 537 cnt.v_pdpages++; 538 if (p->wire_count != 0 || 539 p->hold_count != 0 || 540 p->busy != 0 || 541 (p->flags & (PG_BUSY|PG_UNMANAGED)) || 542 !pmap_page_exists_quick(pmap, p)) { 543 p = next; 544 continue; 545 } 546 actcount = pmap_ts_referenced(p); 547 if (actcount) { 548 vm_page_flag_set(p, PG_REFERENCED); 549 } else if (p->flags & PG_REFERENCED) { 550 actcount = 1; 551 } 552 if ((p->queue != PQ_ACTIVE) && 553 (p->flags & PG_REFERENCED)) { 554 vm_page_activate(p); 555 p->act_count += actcount; 556 vm_page_flag_clear(p, PG_REFERENCED); 557 } else if (p->queue == PQ_ACTIVE) { 558 if ((p->flags & PG_REFERENCED) == 0) { 559 p->act_count -= min(p->act_count, ACT_DECLINE); 560 if (!remove_mode && (vm_pageout_algorithm || (p->act_count == 0))) { 561 pmap_remove_all(p); 562 vm_page_deactivate(p); 563 } else { 564 vm_pageq_requeue(p); 565 } 566 } else { 567 vm_page_activate(p); 568 vm_page_flag_clear(p, PG_REFERENCED); 569 if (p->act_count < (ACT_MAX - ACT_ADVANCE)) 570 p->act_count += ACT_ADVANCE; 571 vm_pageq_requeue(p); 572 } 573 } else if (p->queue == PQ_INACTIVE) { 574 pmap_remove_all(p); 575 } 576 p = next; 577 } 578 vm_page_unlock_queues(); 579 if ((backing_object = object->backing_object) == NULL) 580 goto unlock_return; 581 VM_OBJECT_LOCK(backing_object); 582 if (object != first_object) 583 VM_OBJECT_UNLOCK(object); 584 } 585 unlock_return: 586 if (object != first_object) 587 VM_OBJECT_UNLOCK(object); 588 } 589 590 /* 591 * deactivate some number of pages in a map, try to do it fairly, but 592 * that is really hard to do. 593 */ 594 static void 595 vm_pageout_map_deactivate_pages(map, desired) 596 vm_map_t map; 597 long desired; 598 { 599 vm_map_entry_t tmpe; 600 vm_object_t obj, bigobj; 601 int nothingwired; 602 603 if (!vm_map_trylock(map)) 604 return; 605 606 bigobj = NULL; 607 nothingwired = TRUE; 608 609 /* 610 * first, search out the biggest object, and try to free pages from 611 * that. 612 */ 613 tmpe = map->header.next; 614 while (tmpe != &map->header) { 615 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 616 obj = tmpe->object.vm_object; 617 if (obj != NULL && VM_OBJECT_TRYLOCK(obj)) { 618 if (obj->shadow_count <= 1 && 619 (bigobj == NULL || 620 bigobj->resident_page_count < obj->resident_page_count)) { 621 if (bigobj != NULL) 622 VM_OBJECT_UNLOCK(bigobj); 623 bigobj = obj; 624 } else 625 VM_OBJECT_UNLOCK(obj); 626 } 627 } 628 if (tmpe->wired_count > 0) 629 nothingwired = FALSE; 630 tmpe = tmpe->next; 631 } 632 633 if (bigobj != NULL) { 634 vm_pageout_object_deactivate_pages(map->pmap, bigobj, desired); 635 VM_OBJECT_UNLOCK(bigobj); 636 } 637 /* 638 * Next, hunt around for other pages to deactivate. We actually 639 * do this search sort of wrong -- .text first is not the best idea. 640 */ 641 tmpe = map->header.next; 642 while (tmpe != &map->header) { 643 if (pmap_resident_count(vm_map_pmap(map)) <= desired) 644 break; 645 if ((tmpe->eflags & MAP_ENTRY_IS_SUB_MAP) == 0) { 646 obj = tmpe->object.vm_object; 647 if (obj != NULL) { 648 VM_OBJECT_LOCK(obj); 649 vm_pageout_object_deactivate_pages(map->pmap, obj, desired); 650 VM_OBJECT_UNLOCK(obj); 651 } 652 } 653 tmpe = tmpe->next; 654 } 655 656 /* 657 * Remove all mappings if a process is swapped out, this will free page 658 * table pages. 659 */ 660 if (desired == 0 && nothingwired) { 661 pmap_remove(vm_map_pmap(map), vm_map_min(map), 662 vm_map_max(map)); 663 } 664 vm_map_unlock(map); 665 } 666 #endif /* !defined(NO_SWAPPING) */ 667 668 /* 669 * vm_pageout_scan does the dirty work for the pageout daemon. 670 */ 671 static void 672 vm_pageout_scan(int pass) 673 { 674 vm_page_t m, next; 675 struct vm_page marker; 676 int page_shortage, maxscan, pcount; 677 int addl_page_shortage, addl_page_shortage_init; 678 struct proc *p, *bigproc; 679 struct thread *td; 680 vm_offset_t size, bigsize; 681 vm_object_t object; 682 int actcount, cache_cur, cache_first_failure; 683 static int cache_last_free; 684 int vnodes_skipped = 0; 685 int maxlaunder; 686 687 mtx_lock(&Giant); 688 /* 689 * Decrease registered cache sizes. 690 */ 691 EVENTHANDLER_INVOKE(vm_lowmem, 0); 692 /* 693 * We do this explicitly after the caches have been drained above. 694 */ 695 uma_reclaim(); 696 697 addl_page_shortage_init = atomic_readandclear_int(&vm_pageout_deficit); 698 699 /* 700 * Calculate the number of pages we want to either free or move 701 * to the cache. 702 */ 703 page_shortage = vm_paging_target() + addl_page_shortage_init; 704 705 /* 706 * Initialize our marker 707 */ 708 bzero(&marker, sizeof(marker)); 709 marker.flags = PG_BUSY | PG_FICTITIOUS | PG_MARKER; 710 marker.queue = PQ_INACTIVE; 711 marker.wire_count = 1; 712 713 /* 714 * Start scanning the inactive queue for pages we can move to the 715 * cache or free. The scan will stop when the target is reached or 716 * we have scanned the entire inactive queue. Note that m->act_count 717 * is not used to form decisions for the inactive queue, only for the 718 * active queue. 719 * 720 * maxlaunder limits the number of dirty pages we flush per scan. 721 * For most systems a smaller value (16 or 32) is more robust under 722 * extreme memory and disk pressure because any unnecessary writes 723 * to disk can result in extreme performance degredation. However, 724 * systems with excessive dirty pages (especially when MAP_NOSYNC is 725 * used) will die horribly with limited laundering. If the pageout 726 * daemon cannot clean enough pages in the first pass, we let it go 727 * all out in succeeding passes. 728 */ 729 if ((maxlaunder = vm_max_launder) <= 1) 730 maxlaunder = 1; 731 if (pass) 732 maxlaunder = 10000; 733 vm_page_lock_queues(); 734 rescan0: 735 addl_page_shortage = addl_page_shortage_init; 736 maxscan = cnt.v_inactive_count; 737 738 for (m = TAILQ_FIRST(&vm_page_queues[PQ_INACTIVE].pl); 739 m != NULL && maxscan-- > 0 && page_shortage > 0; 740 m = next) { 741 742 cnt.v_pdpages++; 743 744 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE) { 745 goto rescan0; 746 } 747 748 next = TAILQ_NEXT(m, pageq); 749 object = m->object; 750 751 /* 752 * skip marker pages 753 */ 754 if (m->flags & PG_MARKER) 755 continue; 756 757 /* 758 * A held page may be undergoing I/O, so skip it. 759 */ 760 if (m->hold_count) { 761 vm_pageq_requeue(m); 762 addl_page_shortage++; 763 continue; 764 } 765 /* 766 * Don't mess with busy pages, keep in the front of the 767 * queue, most likely are being paged out. 768 */ 769 if (!VM_OBJECT_TRYLOCK(object) && 770 (!vm_pageout_fallback_object_lock(m, &next) || 771 m->hold_count != 0)) { 772 VM_OBJECT_UNLOCK(object); 773 addl_page_shortage++; 774 continue; 775 } 776 if (m->busy || (m->flags & PG_BUSY)) { 777 VM_OBJECT_UNLOCK(object); 778 addl_page_shortage++; 779 continue; 780 } 781 782 /* 783 * If the object is not being used, we ignore previous 784 * references. 785 */ 786 if (object->ref_count == 0) { 787 vm_page_flag_clear(m, PG_REFERENCED); 788 pmap_clear_reference(m); 789 790 /* 791 * Otherwise, if the page has been referenced while in the 792 * inactive queue, we bump the "activation count" upwards, 793 * making it less likely that the page will be added back to 794 * the inactive queue prematurely again. Here we check the 795 * page tables (or emulated bits, if any), given the upper 796 * level VM system not knowing anything about existing 797 * references. 798 */ 799 } else if (((m->flags & PG_REFERENCED) == 0) && 800 (actcount = pmap_ts_referenced(m))) { 801 vm_page_activate(m); 802 VM_OBJECT_UNLOCK(object); 803 m->act_count += (actcount + ACT_ADVANCE); 804 continue; 805 } 806 807 /* 808 * If the upper level VM system knows about any page 809 * references, we activate the page. We also set the 810 * "activation count" higher than normal so that we will less 811 * likely place pages back onto the inactive queue again. 812 */ 813 if ((m->flags & PG_REFERENCED) != 0) { 814 vm_page_flag_clear(m, PG_REFERENCED); 815 actcount = pmap_ts_referenced(m); 816 vm_page_activate(m); 817 VM_OBJECT_UNLOCK(object); 818 m->act_count += (actcount + ACT_ADVANCE + 1); 819 continue; 820 } 821 822 /* 823 * If the upper level VM system doesn't know anything about 824 * the page being dirty, we have to check for it again. As 825 * far as the VM code knows, any partially dirty pages are 826 * fully dirty. 827 */ 828 if (m->dirty == 0 && !pmap_is_modified(m)) { 829 /* 830 * Avoid a race condition: Unless write access is 831 * removed from the page, another processor could 832 * modify it before all access is removed by the call 833 * to vm_page_cache() below. If vm_page_cache() finds 834 * that the page has been modified when it removes all 835 * access, it panics because it cannot cache dirty 836 * pages. In principle, we could eliminate just write 837 * access here rather than all access. In the expected 838 * case, when there are no last instant modifications 839 * to the page, removing all access will be cheaper 840 * overall. 841 */ 842 if ((m->flags & PG_WRITEABLE) != 0) 843 pmap_remove_all(m); 844 } else { 845 vm_page_dirty(m); 846 } 847 848 if (m->valid == 0) { 849 /* 850 * Invalid pages can be easily freed 851 */ 852 vm_page_free(m); 853 cnt.v_dfree++; 854 --page_shortage; 855 } else if (m->dirty == 0) { 856 /* 857 * Clean pages can be placed onto the cache queue. 858 * This effectively frees them. 859 */ 860 vm_page_cache(m); 861 --page_shortage; 862 } else if ((m->flags & PG_WINATCFLS) == 0 && pass == 0) { 863 /* 864 * Dirty pages need to be paged out, but flushing 865 * a page is extremely expensive verses freeing 866 * a clean page. Rather then artificially limiting 867 * the number of pages we can flush, we instead give 868 * dirty pages extra priority on the inactive queue 869 * by forcing them to be cycled through the queue 870 * twice before being flushed, after which the 871 * (now clean) page will cycle through once more 872 * before being freed. This significantly extends 873 * the thrash point for a heavily loaded machine. 874 */ 875 vm_page_flag_set(m, PG_WINATCFLS); 876 vm_pageq_requeue(m); 877 } else if (maxlaunder > 0) { 878 /* 879 * We always want to try to flush some dirty pages if 880 * we encounter them, to keep the system stable. 881 * Normally this number is small, but under extreme 882 * pressure where there are insufficient clean pages 883 * on the inactive queue, we may have to go all out. 884 */ 885 int swap_pageouts_ok; 886 struct vnode *vp = NULL; 887 struct mount *mp; 888 889 if ((object->type != OBJT_SWAP) && (object->type != OBJT_DEFAULT)) { 890 swap_pageouts_ok = 1; 891 } else { 892 swap_pageouts_ok = !(defer_swap_pageouts || disable_swap_pageouts); 893 swap_pageouts_ok |= (!disable_swap_pageouts && defer_swap_pageouts && 894 vm_page_count_min()); 895 896 } 897 898 /* 899 * We don't bother paging objects that are "dead". 900 * Those objects are in a "rundown" state. 901 */ 902 if (!swap_pageouts_ok || (object->flags & OBJ_DEAD)) { 903 VM_OBJECT_UNLOCK(object); 904 vm_pageq_requeue(m); 905 continue; 906 } 907 908 /* 909 * Following operations may unlock 910 * vm_page_queue_mtx, invalidating the 'next' 911 * pointer. To prevent an inordinate number 912 * of restarts we use our marker to remember 913 * our place. 914 * 915 */ 916 TAILQ_INSERT_AFTER(&vm_page_queues[PQ_INACTIVE].pl, 917 m, &marker, pageq); 918 /* 919 * The object is already known NOT to be dead. It 920 * is possible for the vget() to block the whole 921 * pageout daemon, but the new low-memory handling 922 * code should prevent it. 923 * 924 * The previous code skipped locked vnodes and, worse, 925 * reordered pages in the queue. This results in 926 * completely non-deterministic operation and, on a 927 * busy system, can lead to extremely non-optimal 928 * pageouts. For example, it can cause clean pages 929 * to be freed and dirty pages to be moved to the end 930 * of the queue. Since dirty pages are also moved to 931 * the end of the queue once-cleaned, this gives 932 * way too large a weighting to defering the freeing 933 * of dirty pages. 934 * 935 * We can't wait forever for the vnode lock, we might 936 * deadlock due to a vn_read() getting stuck in 937 * vm_wait while holding this vnode. We skip the 938 * vnode if we can't get it in a reasonable amount 939 * of time. 940 */ 941 if (object->type == OBJT_VNODE) { 942 vp = object->handle; 943 mp = NULL; 944 if (vp->v_type == VREG && 945 vn_start_write(vp, &mp, V_NOWAIT) != 0) { 946 ++pageout_lock_miss; 947 if (object->flags & OBJ_MIGHTBEDIRTY) 948 vnodes_skipped++; 949 vp = NULL; 950 goto unlock_and_continue; 951 } 952 vm_page_unlock_queues(); 953 VI_LOCK(vp); 954 VM_OBJECT_UNLOCK(object); 955 if (vget(vp, LK_EXCLUSIVE | LK_INTERLOCK | 956 LK_TIMELOCK, curthread)) { 957 VM_OBJECT_LOCK(object); 958 vm_page_lock_queues(); 959 ++pageout_lock_miss; 960 vn_finished_write(mp); 961 if (object->flags & OBJ_MIGHTBEDIRTY) 962 vnodes_skipped++; 963 vp = NULL; 964 goto unlock_and_continue; 965 } 966 VM_OBJECT_LOCK(object); 967 vm_page_lock_queues(); 968 /* 969 * The page might have been moved to another 970 * queue during potential blocking in vget() 971 * above. The page might have been freed and 972 * reused for another vnode. The object might 973 * have been reused for another vnode. 974 */ 975 if (VM_PAGE_GETQUEUE(m) != PQ_INACTIVE || 976 m->object != object || 977 object->handle != vp || 978 TAILQ_NEXT(m, pageq) != &marker) { 979 if (object->flags & OBJ_MIGHTBEDIRTY) 980 vnodes_skipped++; 981 goto unlock_and_continue; 982 } 983 984 /* 985 * The page may have been busied during the 986 * blocking in vput(); We don't move the 987 * page back onto the end of the queue so that 988 * statistics are more correct if we don't. 989 */ 990 if (m->busy || (m->flags & PG_BUSY)) { 991 goto unlock_and_continue; 992 } 993 994 /* 995 * If the page has become held it might 996 * be undergoing I/O, so skip it 997 */ 998 if (m->hold_count) { 999 vm_pageq_requeue(m); 1000 if (object->flags & OBJ_MIGHTBEDIRTY) 1001 vnodes_skipped++; 1002 goto unlock_and_continue; 1003 } 1004 } 1005 1006 /* 1007 * If a page is dirty, then it is either being washed 1008 * (but not yet cleaned) or it is still in the 1009 * laundry. If it is still in the laundry, then we 1010 * start the cleaning operation. 1011 * 1012 * decrement page_shortage on success to account for 1013 * the (future) cleaned page. Otherwise we could wind 1014 * up laundering or cleaning too many pages. 1015 */ 1016 if (vm_pageout_clean(m) != 0) { 1017 --page_shortage; 1018 --maxlaunder; 1019 } 1020 unlock_and_continue: 1021 VM_OBJECT_UNLOCK(object); 1022 if (vp) { 1023 vm_page_unlock_queues(); 1024 vput(vp); 1025 vn_finished_write(mp); 1026 vm_page_lock_queues(); 1027 } 1028 next = TAILQ_NEXT(&marker, pageq); 1029 TAILQ_REMOVE(&vm_page_queues[PQ_INACTIVE].pl, 1030 &marker, pageq); 1031 continue; 1032 } 1033 VM_OBJECT_UNLOCK(object); 1034 } 1035 1036 /* 1037 * Compute the number of pages we want to try to move from the 1038 * active queue to the inactive queue. 1039 */ 1040 page_shortage = vm_paging_target() + 1041 cnt.v_inactive_target - cnt.v_inactive_count; 1042 page_shortage += addl_page_shortage; 1043 1044 /* 1045 * Scan the active queue for things we can deactivate. We nominally 1046 * track the per-page activity counter and use it to locate 1047 * deactivation candidates. 1048 */ 1049 pcount = cnt.v_active_count; 1050 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1051 1052 while ((m != NULL) && (pcount-- > 0) && (page_shortage > 0)) { 1053 1054 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1055 ("vm_pageout_scan: page %p isn't active", m)); 1056 1057 next = TAILQ_NEXT(m, pageq); 1058 object = m->object; 1059 if ((m->flags & PG_MARKER) != 0) { 1060 m = next; 1061 continue; 1062 } 1063 if (!VM_OBJECT_TRYLOCK(object) && 1064 !vm_pageout_fallback_object_lock(m, &next)) { 1065 VM_OBJECT_UNLOCK(object); 1066 m = next; 1067 continue; 1068 } 1069 1070 /* 1071 * Don't deactivate pages that are busy. 1072 */ 1073 if ((m->busy != 0) || 1074 (m->flags & PG_BUSY) || 1075 (m->hold_count != 0)) { 1076 VM_OBJECT_UNLOCK(object); 1077 vm_pageq_requeue(m); 1078 m = next; 1079 continue; 1080 } 1081 1082 /* 1083 * The count for pagedaemon pages is done after checking the 1084 * page for eligibility... 1085 */ 1086 cnt.v_pdpages++; 1087 1088 /* 1089 * Check to see "how much" the page has been used. 1090 */ 1091 actcount = 0; 1092 if (object->ref_count != 0) { 1093 if (m->flags & PG_REFERENCED) { 1094 actcount += 1; 1095 } 1096 actcount += pmap_ts_referenced(m); 1097 if (actcount) { 1098 m->act_count += ACT_ADVANCE + actcount; 1099 if (m->act_count > ACT_MAX) 1100 m->act_count = ACT_MAX; 1101 } 1102 } 1103 1104 /* 1105 * Since we have "tested" this bit, we need to clear it now. 1106 */ 1107 vm_page_flag_clear(m, PG_REFERENCED); 1108 1109 /* 1110 * Only if an object is currently being used, do we use the 1111 * page activation count stats. 1112 */ 1113 if (actcount && (object->ref_count != 0)) { 1114 vm_pageq_requeue(m); 1115 } else { 1116 m->act_count -= min(m->act_count, ACT_DECLINE); 1117 if (vm_pageout_algorithm || 1118 object->ref_count == 0 || 1119 m->act_count == 0) { 1120 page_shortage--; 1121 if (object->ref_count == 0) { 1122 pmap_remove_all(m); 1123 if (m->dirty == 0) 1124 vm_page_cache(m); 1125 else 1126 vm_page_deactivate(m); 1127 } else { 1128 vm_page_deactivate(m); 1129 } 1130 } else { 1131 vm_pageq_requeue(m); 1132 } 1133 } 1134 VM_OBJECT_UNLOCK(object); 1135 m = next; 1136 } 1137 1138 /* 1139 * We try to maintain some *really* free pages, this allows interrupt 1140 * code to be guaranteed space. Since both cache and free queues 1141 * are considered basically 'free', moving pages from cache to free 1142 * does not effect other calculations. 1143 */ 1144 cache_cur = cache_last_free; 1145 cache_first_failure = -1; 1146 while (cnt.v_free_count < cnt.v_free_reserved && (cache_cur = 1147 (cache_cur + PQ_PRIME2) & PQ_COLORMASK) != cache_first_failure) { 1148 TAILQ_FOREACH(m, &vm_page_queues[PQ_CACHE + cache_cur].pl, 1149 pageq) { 1150 KASSERT(m->dirty == 0, 1151 ("Found dirty cache page %p", m)); 1152 KASSERT(!pmap_page_is_mapped(m), 1153 ("Found mapped cache page %p", m)); 1154 KASSERT((m->flags & PG_UNMANAGED) == 0, 1155 ("Found unmanaged cache page %p", m)); 1156 KASSERT(m->wire_count == 0, 1157 ("Found wired cache page %p", m)); 1158 if (m->hold_count == 0 && VM_OBJECT_TRYLOCK(object = 1159 m->object)) { 1160 KASSERT((m->flags & PG_BUSY) == 0 && 1161 m->busy == 0, ("Found busy cache page %p", 1162 m)); 1163 vm_page_free(m); 1164 VM_OBJECT_UNLOCK(object); 1165 cnt.v_dfree++; 1166 cache_last_free = cache_cur; 1167 cache_first_failure = -1; 1168 break; 1169 } 1170 } 1171 if (m == NULL && cache_first_failure == -1) 1172 cache_first_failure = cache_cur; 1173 } 1174 vm_page_unlock_queues(); 1175 #if !defined(NO_SWAPPING) 1176 /* 1177 * Idle process swapout -- run once per second. 1178 */ 1179 if (vm_swap_idle_enabled) { 1180 static long lsec; 1181 if (time_second != lsec) { 1182 vm_pageout_req_swapout |= VM_SWAP_IDLE; 1183 vm_req_vmdaemon(); 1184 lsec = time_second; 1185 } 1186 } 1187 #endif 1188 1189 /* 1190 * If we didn't get enough free pages, and we have skipped a vnode 1191 * in a writeable object, wakeup the sync daemon. And kick swapout 1192 * if we did not get enough free pages. 1193 */ 1194 if (vm_paging_target() > 0) { 1195 if (vnodes_skipped && vm_page_count_min()) 1196 (void) speedup_syncer(); 1197 #if !defined(NO_SWAPPING) 1198 if (vm_swap_enabled && vm_page_count_target()) { 1199 vm_req_vmdaemon(); 1200 vm_pageout_req_swapout |= VM_SWAP_NORMAL; 1201 } 1202 #endif 1203 } 1204 1205 /* 1206 * If we are critically low on one of RAM or swap and low on 1207 * the other, kill the largest process. However, we avoid 1208 * doing this on the first pass in order to give ourselves a 1209 * chance to flush out dirty vnode-backed pages and to allow 1210 * active pages to be moved to the inactive queue and reclaimed. 1211 * 1212 * We keep the process bigproc locked once we find it to keep anyone 1213 * from messing with it; however, there is a possibility of 1214 * deadlock if process B is bigproc and one of it's child processes 1215 * attempts to propagate a signal to B while we are waiting for A's 1216 * lock while walking this list. To avoid this, we don't block on 1217 * the process lock but just skip a process if it is already locked. 1218 */ 1219 if (pass != 0 && 1220 ((swap_pager_avail < 64 && vm_page_count_min()) || 1221 (swap_pager_full && vm_paging_target() > 0))) { 1222 bigproc = NULL; 1223 bigsize = 0; 1224 sx_slock(&allproc_lock); 1225 FOREACH_PROC_IN_SYSTEM(p) { 1226 int breakout; 1227 1228 if (PROC_TRYLOCK(p) == 0) 1229 continue; 1230 /* 1231 * If this is a system or protected process, skip it. 1232 */ 1233 if ((p->p_flag & P_SYSTEM) || (p->p_pid == 1) || 1234 (p->p_flag & P_PROTECTED) || 1235 ((p->p_pid < 48) && (swap_pager_avail != 0))) { 1236 PROC_UNLOCK(p); 1237 continue; 1238 } 1239 /* 1240 * If the process is in a non-running type state, 1241 * don't touch it. Check all the threads individually. 1242 */ 1243 mtx_lock_spin(&sched_lock); 1244 breakout = 0; 1245 FOREACH_THREAD_IN_PROC(p, td) { 1246 if (!TD_ON_RUNQ(td) && 1247 !TD_IS_RUNNING(td) && 1248 !TD_IS_SLEEPING(td)) { 1249 breakout = 1; 1250 break; 1251 } 1252 } 1253 if (breakout) { 1254 mtx_unlock_spin(&sched_lock); 1255 PROC_UNLOCK(p); 1256 continue; 1257 } 1258 mtx_unlock_spin(&sched_lock); 1259 /* 1260 * get the process size 1261 */ 1262 if (!vm_map_trylock_read(&p->p_vmspace->vm_map)) { 1263 PROC_UNLOCK(p); 1264 continue; 1265 } 1266 size = vmspace_swap_count(p->p_vmspace); 1267 vm_map_unlock_read(&p->p_vmspace->vm_map); 1268 size += vmspace_resident_count(p->p_vmspace); 1269 /* 1270 * if the this process is bigger than the biggest one 1271 * remember it. 1272 */ 1273 if (size > bigsize) { 1274 if (bigproc != NULL) 1275 PROC_UNLOCK(bigproc); 1276 bigproc = p; 1277 bigsize = size; 1278 } else 1279 PROC_UNLOCK(p); 1280 } 1281 sx_sunlock(&allproc_lock); 1282 if (bigproc != NULL) { 1283 killproc(bigproc, "out of swap space"); 1284 mtx_lock_spin(&sched_lock); 1285 sched_nice(bigproc, PRIO_MIN); 1286 mtx_unlock_spin(&sched_lock); 1287 PROC_UNLOCK(bigproc); 1288 wakeup(&cnt.v_free_count); 1289 } 1290 } 1291 mtx_unlock(&Giant); 1292 } 1293 1294 /* 1295 * This routine tries to maintain the pseudo LRU active queue, 1296 * so that during long periods of time where there is no paging, 1297 * that some statistic accumulation still occurs. This code 1298 * helps the situation where paging just starts to occur. 1299 */ 1300 static void 1301 vm_pageout_page_stats() 1302 { 1303 vm_object_t object; 1304 vm_page_t m,next; 1305 int pcount,tpcount; /* Number of pages to check */ 1306 static int fullintervalcount = 0; 1307 int page_shortage; 1308 1309 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1310 page_shortage = 1311 (cnt.v_inactive_target + cnt.v_cache_max + cnt.v_free_min) - 1312 (cnt.v_free_count + cnt.v_inactive_count + cnt.v_cache_count); 1313 1314 if (page_shortage <= 0) 1315 return; 1316 1317 pcount = cnt.v_active_count; 1318 fullintervalcount += vm_pageout_stats_interval; 1319 if (fullintervalcount < vm_pageout_full_stats_interval) { 1320 tpcount = (vm_pageout_stats_max * cnt.v_active_count) / cnt.v_page_count; 1321 if (pcount > tpcount) 1322 pcount = tpcount; 1323 } else { 1324 fullintervalcount = 0; 1325 } 1326 1327 m = TAILQ_FIRST(&vm_page_queues[PQ_ACTIVE].pl); 1328 while ((m != NULL) && (pcount-- > 0)) { 1329 int actcount; 1330 1331 KASSERT(VM_PAGE_INQUEUE2(m, PQ_ACTIVE), 1332 ("vm_pageout_page_stats: page %p isn't active", m)); 1333 1334 next = TAILQ_NEXT(m, pageq); 1335 object = m->object; 1336 1337 if ((m->flags & PG_MARKER) != 0) { 1338 m = next; 1339 continue; 1340 } 1341 if (!VM_OBJECT_TRYLOCK(object) && 1342 !vm_pageout_fallback_object_lock(m, &next)) { 1343 VM_OBJECT_UNLOCK(object); 1344 m = next; 1345 continue; 1346 } 1347 1348 /* 1349 * Don't deactivate pages that are busy. 1350 */ 1351 if ((m->busy != 0) || 1352 (m->flags & PG_BUSY) || 1353 (m->hold_count != 0)) { 1354 VM_OBJECT_UNLOCK(object); 1355 vm_pageq_requeue(m); 1356 m = next; 1357 continue; 1358 } 1359 1360 actcount = 0; 1361 if (m->flags & PG_REFERENCED) { 1362 vm_page_flag_clear(m, PG_REFERENCED); 1363 actcount += 1; 1364 } 1365 1366 actcount += pmap_ts_referenced(m); 1367 if (actcount) { 1368 m->act_count += ACT_ADVANCE + actcount; 1369 if (m->act_count > ACT_MAX) 1370 m->act_count = ACT_MAX; 1371 vm_pageq_requeue(m); 1372 } else { 1373 if (m->act_count == 0) { 1374 /* 1375 * We turn off page access, so that we have 1376 * more accurate RSS stats. We don't do this 1377 * in the normal page deactivation when the 1378 * system is loaded VM wise, because the 1379 * cost of the large number of page protect 1380 * operations would be higher than the value 1381 * of doing the operation. 1382 */ 1383 pmap_remove_all(m); 1384 vm_page_deactivate(m); 1385 } else { 1386 m->act_count -= min(m->act_count, ACT_DECLINE); 1387 vm_pageq_requeue(m); 1388 } 1389 } 1390 VM_OBJECT_UNLOCK(object); 1391 m = next; 1392 } 1393 } 1394 1395 /* 1396 * vm_pageout is the high level pageout daemon. 1397 */ 1398 static void 1399 vm_pageout() 1400 { 1401 int error, pass; 1402 1403 /* 1404 * Initialize some paging parameters. 1405 */ 1406 cnt.v_interrupt_free_min = 2; 1407 if (cnt.v_page_count < 2000) 1408 vm_pageout_page_count = 8; 1409 1410 /* 1411 * v_free_reserved needs to include enough for the largest 1412 * swap pager structures plus enough for any pv_entry structs 1413 * when paging. 1414 */ 1415 if (cnt.v_page_count > 1024) 1416 cnt.v_free_min = 4 + (cnt.v_page_count - 1024) / 200; 1417 else 1418 cnt.v_free_min = 4; 1419 cnt.v_pageout_free_min = (2*MAXBSIZE)/PAGE_SIZE + 1420 cnt.v_interrupt_free_min; 1421 cnt.v_free_reserved = vm_pageout_page_count + 1422 cnt.v_pageout_free_min + (cnt.v_page_count / 768) + PQ_NUMCOLORS; 1423 cnt.v_free_severe = cnt.v_free_min / 2; 1424 cnt.v_free_min += cnt.v_free_reserved; 1425 cnt.v_free_severe += cnt.v_free_reserved; 1426 1427 /* 1428 * v_free_target and v_cache_min control pageout hysteresis. Note 1429 * that these are more a measure of the VM cache queue hysteresis 1430 * then the VM free queue. Specifically, v_free_target is the 1431 * high water mark (free+cache pages). 1432 * 1433 * v_free_reserved + v_cache_min (mostly means v_cache_min) is the 1434 * low water mark, while v_free_min is the stop. v_cache_min must 1435 * be big enough to handle memory needs while the pageout daemon 1436 * is signalled and run to free more pages. 1437 */ 1438 if (cnt.v_free_count > 6144) 1439 cnt.v_free_target = 4 * cnt.v_free_min + cnt.v_free_reserved; 1440 else 1441 cnt.v_free_target = 2 * cnt.v_free_min + cnt.v_free_reserved; 1442 1443 if (cnt.v_free_count > 2048) { 1444 cnt.v_cache_min = cnt.v_free_target; 1445 cnt.v_cache_max = 2 * cnt.v_cache_min; 1446 cnt.v_inactive_target = (3 * cnt.v_free_target) / 2; 1447 } else { 1448 cnt.v_cache_min = 0; 1449 cnt.v_cache_max = 0; 1450 cnt.v_inactive_target = cnt.v_free_count / 4; 1451 } 1452 if (cnt.v_inactive_target > cnt.v_free_count / 3) 1453 cnt.v_inactive_target = cnt.v_free_count / 3; 1454 1455 /* XXX does not really belong here */ 1456 if (vm_page_max_wired == 0) 1457 vm_page_max_wired = cnt.v_free_count / 3; 1458 1459 if (vm_pageout_stats_max == 0) 1460 vm_pageout_stats_max = cnt.v_free_target; 1461 1462 /* 1463 * Set interval in seconds for stats scan. 1464 */ 1465 if (vm_pageout_stats_interval == 0) 1466 vm_pageout_stats_interval = 5; 1467 if (vm_pageout_full_stats_interval == 0) 1468 vm_pageout_full_stats_interval = vm_pageout_stats_interval * 4; 1469 1470 swap_pager_swap_init(); 1471 pass = 0; 1472 /* 1473 * The pageout daemon is never done, so loop forever. 1474 */ 1475 while (TRUE) { 1476 vm_page_lock_queues(); 1477 /* 1478 * If we have enough free memory, wakeup waiters. Do 1479 * not clear vm_pages_needed until we reach our target, 1480 * otherwise we may be woken up over and over again and 1481 * waste a lot of cpu. 1482 */ 1483 if (vm_pages_needed && !vm_page_count_min()) { 1484 if (!vm_paging_needed()) 1485 vm_pages_needed = 0; 1486 wakeup(&cnt.v_free_count); 1487 } 1488 if (vm_pages_needed) { 1489 /* 1490 * Still not done, take a second pass without waiting 1491 * (unlimited dirty cleaning), otherwise sleep a bit 1492 * and try again. 1493 */ 1494 ++pass; 1495 if (pass > 1) 1496 msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1497 "psleep", hz/2); 1498 } else { 1499 /* 1500 * Good enough, sleep & handle stats. Prime the pass 1501 * for the next run. 1502 */ 1503 if (pass > 1) 1504 pass = 1; 1505 else 1506 pass = 0; 1507 error = msleep(&vm_pages_needed, &vm_page_queue_mtx, PVM, 1508 "psleep", vm_pageout_stats_interval * hz); 1509 if (error && !vm_pages_needed) { 1510 pass = 0; 1511 vm_pageout_page_stats(); 1512 vm_page_unlock_queues(); 1513 continue; 1514 } 1515 } 1516 if (vm_pages_needed) 1517 cnt.v_pdwakeups++; 1518 vm_page_unlock_queues(); 1519 vm_pageout_scan(pass); 1520 } 1521 } 1522 1523 /* 1524 * Unless the page queue lock is held by the caller, this function 1525 * should be regarded as advisory. Specifically, the caller should 1526 * not msleep() on &cnt.v_free_count following this function unless 1527 * the page queue lock is held until the msleep() is performed. 1528 */ 1529 void 1530 pagedaemon_wakeup() 1531 { 1532 1533 if (!vm_pages_needed && curthread->td_proc != pageproc) { 1534 vm_pages_needed = 1; 1535 wakeup(&vm_pages_needed); 1536 } 1537 } 1538 1539 #if !defined(NO_SWAPPING) 1540 static void 1541 vm_req_vmdaemon() 1542 { 1543 static int lastrun = 0; 1544 1545 if ((ticks > (lastrun + hz)) || (ticks < lastrun)) { 1546 wakeup(&vm_daemon_needed); 1547 lastrun = ticks; 1548 } 1549 } 1550 1551 static void 1552 vm_daemon() 1553 { 1554 struct rlimit rsslim; 1555 struct proc *p; 1556 struct thread *td; 1557 int breakout; 1558 1559 mtx_lock(&Giant); 1560 while (TRUE) { 1561 tsleep(&vm_daemon_needed, PPAUSE, "psleep", 0); 1562 if (vm_pageout_req_swapout) { 1563 swapout_procs(vm_pageout_req_swapout); 1564 vm_pageout_req_swapout = 0; 1565 } 1566 /* 1567 * scan the processes for exceeding their rlimits or if 1568 * process is swapped out -- deactivate pages 1569 */ 1570 sx_slock(&allproc_lock); 1571 LIST_FOREACH(p, &allproc, p_list) { 1572 vm_pindex_t limit, size; 1573 1574 /* 1575 * if this is a system process or if we have already 1576 * looked at this process, skip it. 1577 */ 1578 PROC_LOCK(p); 1579 if (p->p_flag & (P_SYSTEM | P_WEXIT)) { 1580 PROC_UNLOCK(p); 1581 continue; 1582 } 1583 /* 1584 * if the process is in a non-running type state, 1585 * don't touch it. 1586 */ 1587 mtx_lock_spin(&sched_lock); 1588 breakout = 0; 1589 FOREACH_THREAD_IN_PROC(p, td) { 1590 if (!TD_ON_RUNQ(td) && 1591 !TD_IS_RUNNING(td) && 1592 !TD_IS_SLEEPING(td)) { 1593 breakout = 1; 1594 break; 1595 } 1596 } 1597 mtx_unlock_spin(&sched_lock); 1598 if (breakout) { 1599 PROC_UNLOCK(p); 1600 continue; 1601 } 1602 /* 1603 * get a limit 1604 */ 1605 lim_rlimit(p, RLIMIT_RSS, &rsslim); 1606 limit = OFF_TO_IDX( 1607 qmin(rsslim.rlim_cur, rsslim.rlim_max)); 1608 1609 /* 1610 * let processes that are swapped out really be 1611 * swapped out set the limit to nothing (will force a 1612 * swap-out.) 1613 */ 1614 if ((p->p_sflag & PS_INMEM) == 0) 1615 limit = 0; /* XXX */ 1616 PROC_UNLOCK(p); 1617 1618 size = vmspace_resident_count(p->p_vmspace); 1619 if (limit >= 0 && size >= limit) { 1620 vm_pageout_map_deactivate_pages( 1621 &p->p_vmspace->vm_map, limit); 1622 } 1623 } 1624 sx_sunlock(&allproc_lock); 1625 } 1626 } 1627 #endif /* !defined(NO_SWAPPING) */ 1628